A credit and counterparty team reviewed companies on a fixed checklist. The process was consistent in principle and slow in practice, and it treated a thin-margin retailer the same way as a capital-intensive manufacturer. Analysts spent the same time on healthy companies as on the ones that were about to fail, because nothing in the process told them where to look first.
The decision
The team did not need a verdict on whether each company would fail. It needed to decide which companies an analyst should open this week, in what order, and what the analyst should check first. With limited review capacity and a bankruptcy rate in the low single digits, the useful output was a defensible ranking with an explanation — not a single accuracy score.
The cost asymmetry also shaped the work. A missed bankruptcy is expensive; an unnecessary review costs analyst time. The threshold had to reflect that difference, not default to one half.
What we had to work with
Published financial ratios covered the usual ground:
- Profitability: margins, return on assets and return on equity.
- Liquidity: current and quick ratios, working capital and cash position.
- Leverage and solvency: debt ratios, interest coverage and equity structure.
- Activity and cash flow: turnover, asset efficiency and operating cash flow relative to liabilities.
The outcome was a binary bankruptcy flag, heavily imbalanced. Many ratios were correlated with each other, several were near-constant, and a small number of extreme observations dominated any model that accepted them at face value.
Three problems the checklist could not see
- Imbalance. A model that predicts "no bankruptcy" for everyone is correct most of the time and useless every time it matters. Accuracy had to leave the conversation.
- Duplicated information. A dozen liquidity ratios are not a dozen independent facts. Left unreduced, they multiply noise and make the model impossible to explain to a credit committee.
- Time drift. A random split lets the model learn from a company's later years to predict its earlier ones. Financial conditions change, so validation had to move forward in time, not sideways.
How the triage score works
- Audit the ratio set before modelling. Constant variables, duplicated measures and implausible ranges were removed or corrected. A model built on contaminated ratios learns the contamination.
- Reduce deliberately. Univariate screening and correlation pruning produced a smaller, interpretable candidate set. In credit work, a 90-variable black box is not a competitive advantage; it is a review problem.
- Start with a transparent baseline. Logistic regression with binned, monotonic features gave the committee something it could read, and gave the business a natural source of reason codes.
- Make the challenger earn its place. A gradient-boosted model was compared on AUC, KS, lift and calibration. It was adopted only where it improved the decision without becoming unexplainable.
- Weight the rare class against the real cost. Class weighting and threshold selection reflected the relative cost of a missed bankruptcy versus an unnecessary review. The cut-off became a policy decision with an owner, not a technical default.
- Validate over time and across segments. We tested on later periods and checked results by company size and sector where the data allowed. A score that works only on the largest manufacturers is not a portfolio tool.
- Package for the committee, not the notebook. The output was a risk band, a score, the top contributing ratios and a monitoring plan for population drift and changes in ratio definitions.
What this changes
- Analysts got a consistent first pass. Every company received the same initial screen, and analyst time concentrated on the flagged tail.
- Review became a conversation about causes. Reason codes turned a ranking into a discussion about which ratios drove the flag and whether the underlying statement item was reliable.
- The committee could see the policy. Bands, thresholds and cost assumptions were documented, so approvals and overrides could be reviewed rather than reconstructed from memory.
What we delivered
- A triage score with risk bands and analyst-facing reason codes.
- A validation pack comparing the interpretable baseline with the challenger model.
- A review-capacity policy mapping score bands to analyst actions.
- Documentation suitable for credit-committee and internal-audit review.
Where this applies
This fits banks, insurers, leasing companies and any team that reviews counterparties from financial statements. The minimum inputs are company identifiers, at least three years of comparable statements, a bankruptcy or distress outcome and a known review capacity. The first deliverable is usually a ratio audit and a time-aware validation design — the two things that most often change the answer before a model is chosen.
This is an illustrative case study built on an anonymized dataset; it does not report results from a specific client engagement. It reflects our standard method for financial-distress screening.
